Seong-Won Choi, Myung-Il Roh, In-Chang Yeo, "A Method for Ship Collision Avoidance Based on Deep Reinforcement Learning Considering Uncertainty", Proceedings of OMAE(International Conference on Ocean, Offshore and Arctic Engineering) 2026, Tokyo, Japan, 2026.06.07-12
International Conference
2025.11.18 12:57
Seong-Won Choi, Myung-Il Roh, In-Chang Yeo, "A Method for Ship Collision Avoidance Based on Deep Reinforcement Learning Considering Uncertainty", Proceedings of OMAE 2026, Tokyo, Japan, 2026.06.07-12
조회 수 548
| 첨부 '1' |
|---|
| Abstract | In complex maritime environments, various types of ships coexist, resulting in overlapping routes and unpredictable movements. Such complexity can often lead to serious maritime accidents, including collisions and groundings. To ensure safe navigation, it is essential to accurately perceive the dynamic surroundings of the own ship and secure a safe route through appropriate avoidance maneuvers. Traditional collision avoidance methods, such as the VO (Velocity Obstacles) method, are advantageous for calculating real-time avoidance maneuvers by considering the potential collision risk with surrounding TSs (target ships). However, traditional methods for collision avoidance degrade significantly when the positional or velocity information of the TSs is inaccurate. In complex maritime environments, these methods may generate overly conservative routes with limited maneuverability. Moreover, maritime regulations such as the COLREGs (International Regulations for Preventing Collisions at Sea) often contain ambiguous provisions, making it difficult to integrate them effectively into traditional methods for collision avoidance. To address these limitations, we propose a method for ship collision avoidance based on DRL (Deep Reinforcement Learning) that accounts for uncertainty. In this method, the SAC (Soft Actor-Critic) algorithm, the DRL framework is known for its stability and sample efficiency. The proposed method introduces an improved collision risk assessment method that uses the AR (Approach Rate) to calculate collision risk more smoothly and reliably than traditional CPA (Closest Point of Approach) methods. Furthermore, an attention mechanism is incorporated to effectively consider information from multiple TSs, enabling the model to prioritize critical ships based on their relative importance during avoidance decisions. To ensure compliance with real-world navigation rules, the reward function is designed to reflect COLREGs-based constraints, encouraging rule-consistent avoidance maneuvers. During verification, uncertainties in TS positions, speeds, and headings are modeled using a bivariate normal distribution, allowing performance evaluation under realistic sensor noise. Simulation experiments demonstrate that the proposed method achieves safer, more robust collision avoidance maneuvers than traditional methods, maintaining reliability even in uncertain, dynamic maritime environments. |
|---|---|
| Publication Date | 2026-06-09 |
